Feature-based image registration algorithm for image stitching applications on mobile devices

In this paper, we propose a new feature-based image registration algorithm considering the mobile devices with limited computing power. The algorithm is consisted of two steps: first we reject severe outliers in putative correspondences by considering a similarity transform as a motion model between two views. This approach can find most of the inliers very efficiently, but some non-severe outliers still remain due to its coarse approximation. Therefore, in the second step, we apply an optimization method based on the conjugate gradient algorithm to refine a geometric relation between two views and a set of inliers simultaneously. The proposed method is implemented on a mobile device and experimental results have shown that our method is very efficient and yields good registration performance.